Indicadores de Gestión

Selected ground reference points (Section 3.2) were used to validate the classification, both during the classification stage (in conjunction with visual checks) (Section 3.2), and after the final LCM2007 data set was created. This section reports the results of the comparison between the 9127 LCM2007 ground reference polygons and the final LCM2007 product (Table 3.7).

The field trip points were used to identify suitable training areas for the classification (i.e. a training data set), and also to identify suitable points for a ground reference data set against which to validate the product (i.e. a testing data set). The training and testing data set are separate groups of ground reference points. The accuracy of LCM2007 is defined by comparison against the ground reference (testing) data set. In some cases ground reference points and polygons were poorly aligned and in other cases the ground reference points were only appropriate for part of the polygon - this mainly occurred in natural or semi-natural areas, although it did affect some woodlands and fields. Ground reference points were only assigned as validation points if the person conducting the classification was confident that the ground reference point was appropriate for the polygon. This excludes polygons with mixed Broad Habitats (shown by multiple ground reference points within a single-polygon) or where ground reference points fell in adjacent polygons. Once a ground reference point has been accepted as appropriate, for a polygon in the LCM2007 project database, the polygon becomes a ground reference polygon.

Results

The correspondence, between LCM2007 and the 9127 ground reference polygons (Table 3.7), shows that the overall accuracy of LCM2007 is 83%. Information about the classification of individual classes is also given in the correspondence matrix (Table 3.7).

Table 3.7 is a correspondence matrix and is data rich, but due to the number of classes in LCM2007 it is not straightforward to interpret. Using „Bog‟ as an example, the Producer‟s accuracy (see glossary) quantifies how well areas mapped as „Bog‟ in LCM2007 match the ground reference polygons. Table 3.7 shows „Bog‟ is 93% accurate against the ground reference polygons. The User‟s accuracy (see glossary) gives the probability of a parcel of a given class being correctly classified. In the case of „Improved Grassland‟, the User‟s accuracy is 83% and the Producer‟s accuracy is 89%. Therefore, based on the User‟s accuracy of 83%, there is a 0.83 probability that a parcel classified as „Improved Grassland‟ will be correctly classified. The Producer‟s accuracy quantifies how well areas mapped as „Improved Grassland‟ in

44

LCM2007 match the ground reference polygons. The User‟s and Producer‟s accuracy can vary independently of each other, although accurate classes will have high values for both the User‟s and Producer‟s accuracies.

The LCM2007 classes are grouped below based on the User‟s and Producer‟s accuracy. Note that requiring both accuracy values to be above a threshold means that the classes are grouped by the lowest of their two accuracy values, so the location of „Bog‟ is determined by the User‟s accuracy (39%), rather than the Producer‟s accuracy (93%).

 > 90% for User’s accuracy and Producer’s accuracy:

 „Coniferous Woodland‟  „Arable and Horticulture‟  ‘Littoral Rock‟

 > 80% for User’s accuracy and Producer’s accuracy

 Broadleaf woodland  „Improved Grassland‟  Littoral Sediment  Urban  Suburban  Saltmarsh  Freshwater  ‘Supra-littoral Rock‟

 > 70% for User’s accuracy and Producer’s accuracy

 „Fen, Marsh and Swamp‟  Heather

 „Supra-littoral Sediment‟

 > 60% for User’s accuracy and Producer’s accuracy

 Saltwater

 > 50% for User’s accuracy and Producer’s accuracy

 „Calcareous Grassland‟  Heather grassland

< 50% for User’s accuracy and Producer’s accuracy

 Rough grassland  „Acid Grassland‟  „Neutral Grassland‟  „Bog‟

 „Montane Habitats‟

Accuracy is class-specific, with classes using external data sets for knowledge- based enhancements tending to have the lowest accuracy. The problem with „Calcareous Grassland‟ and „Neutral Grassland‟ is largely due to spectral confusion with „Improved Grassland‟ and Rough grassland. The confusion with Rough

grassland is partly due to differences between what it is possible to discern in the

field and what can be achieved by spectral classification, KBEs and a soil data set. The „Montane Habitats‟ designation is expected to be correct for areas mapped as „Montane Habitats‟, but may underestimate in some areas, especially NW Scotland. In some cases it may be useful for user‟s to apply their own bespoke validation, especially where extensive work is planned based on a county or smaller subset of data (see Appendix 4).

45

Table 3.7. Accuracy of LCM2007 based on 9127 LCM2007 ground reference polygons. Green squares correspond at LCM2007 class-level.

Producer‟s accuracy = percentage of ground reference polygons classified correctly. User‟s accuracy = probability (expressed as a percentage) of a

polygon of a particular class being correctly classified.

46

Chapter 4: Comparison with Countryside

Survey in 2007

4.1 Introduction

This chapter presents the results of a comparison of LCM2007 with Countryside Survey (CS) in 2007. The aim is to establish confidence in LCM2007 and quantify the correspondence between the LCM2007 and CS by comparing LCM2007 with:

The 591 Countryside Survey 1x1km squares in Great Britain

The Countryside Survey estimates of Broad Habitat area for England, Wales, Scotland, Northern Ireland and the UK

The first method produces a correspondence matrix. The second method produces a table showing the LCM2007 estimates of Broad Habitat area with the Broad Habitat estimates of area and confidence limits from the CS in 2007.

4.2 Comparison with Countryside Survey squares

Countryside Survey in 2007 surveyed 591 1x1km squares recording Broad Habitat cover, plus more detailed information (Carey et al., 2008). The LCM2007 and Countryside Survey data sets are very different and to increase compatibility between the two, Countryside Survey polygons below the LCM2007 MMU width and area were excluded, as were the Countryside Survey Mosaic and „Boundary and Linear Features‟ classes, as there is no LCM2007 equivalent. Note, the Countryside Survey Mosaic class is used when the field surveyors encounter a mix of Broad Habitat types where it is not possible to map discrete areas of a single habitat exceeding the CS MMU. The field surveyors do record proportions of the Broad Habitats comprising the Mosaic polygons, but it would not be straightforward to include them in this analysis, so this class was excluded. Other differences, between CS in 2007 and LCM2007, such as LCM2007 mapping „Standing Open Water and Canals‟ and „Rivers and Streams‟ as a composite Freshwater class, were accounted for when the correspondence matrix was created. Equivalent data for Northern Ireland were not available to include in this analysis.

An area-based comparison was conducted between the 591 Countryside Survey squares and the corresponding areas of LCM2007. The correspondence was calculated by comparing the polygons in CS with the same area in LCM2007 and recording the area of Broad Habitat. This was conducted for each square in turn, enabling the correspondence for each square to be calculated, and the squares were then aggregated to enable production of correspondence tables for the: UK, England, Scotland and Wales.

Whilst every effort is taken to ensure the accuracy of CS data there are errors, due to habitat definition issues, as well as individual surveyor‟s interpretations. As a result the comparison between LCM2007 and CS in 2007 is about discerning correspondence and divergence, rather than accuracy. In addition, the timing of field

47

survey and image acquisition may not coincide, allowing changes to occur in the intervening period. Given that Countryside Survey is about 81% repeatable (Norton et al., 2009) and that LCM2007 aims for an accuracy of > 80% at the Broad Habitat level, then combining these two accuracy levels suggests a likely correspondence of about 65% (80% x 81% = 65%). However, this does not take into account the different spatial structures of the two products, which although much closer than they used to be, still differ. This will have a tendency to reduce areal correspondence values.

The correspondences are reported for the UK, plus England, Scotland and Wales, and are calculated for a range of thematic levels:

- The Broad Habitats common to LCM2007 and CS in 2007.

- Aggregate class level (Aggregate classes are defined in Table 2.2, Chapter 2).

- Broad Habitat Association (BHA) level.

Broad Habitat Association (BHA)

The Broad Habitat Association concept provides an additional measure of thematic correspondence, which has some similarities to the Aggregate class thematic level, but is more targeted. A key factor underlying the Broad Habitat Association (BHA) concept is that the land cover maps provide land cover, which may relate uniquely to one Broad Habitat or may have a one-to-many relationship with several Broad Habitats. For example, deciduous woodland uniquely maps to the „Broadleaved, Mixed and Yew Woodland‟ Broad Habitat, whereas rough grassland has a one-to- many relationship with habitat, and may be „Acid Grassland‟, „Calcareous Grassland‟, „Neutral Grassland‟, poor quality „Improved Grassland‟ or even in some cases „Fen, Marsh and Swamp‟ or „Bog‟. These distinctions can be difficult to make from remote sensing and even in the field; they are especially difficult when they occur as mosaics where the dominant (most widespread) land cover may vary depending upon the polygon boundaries.

In this type of situation the identification of „associated‟ habitats which form different Broad Habitats, as defined by Jackson (2000), is useful. The cross-habitat links in Table 4.1 formalises some of the uncertainties between mapping land cover and assigning it to a habitat-based classification [see Appendix 5 for further details]. The similarity of land cover for different BH is most pronounced for grassland and semi- natural upland areas, where mosaic landscapes maybe poorly represented spatially by the scale of the parcel-based structure in LCM2007, so these habitats dominate Table 4.1. In practice this means that as well as squares down the main diagonal of the correspondence matrix being accepted as corresponding directly, some additional cells are considered as corresponding directly. For clarity the cells used to calculate the BHA correspondence are shaded orange in Tables 4.3 & 4.5-4.7.

Broad Habitat Association (BHA): BHA identifies prescribed habitat links (Table

4.1) which form allowable correspondence between CS and LCM2007 classes e.g. „Bog‟ is acceptable for „Dwarf Shrub Heath‟ at the BHA-level.

The purpose of using BHAs is to separate correspondences which are slightly different e.g. „Bog‟ and „Dwarf Shrub Heath‟, compared to those that are very different e.g. „Montane Habitats‟ and „Arable and Horticulture‟.

48

Table 4.1. Summary of correspondences allowed for Broad Habitat Association (whether due to transitional habitats, mosaic habitats, limitations of KBE’s or difference in interpretation between Countryside Survey field survey and LCM).

LCM CS Field Survey

„Bog‟ „Montane Habitats‟

„Dwarf Shrub Heath‟ „Acid Grassland‟

„Montane Habitats‟ „Dwarf Shrub Heath‟

„Acid Grassland‟ „Bog‟

„Dwarf Shrub Heath‟ „Acid Grassland‟

Rough grassland „Acid Grassland‟ „Calcareous Grassland‟

„Neutral Grassland‟ „Improved Grassland‟ „Fen, Marsh and Swamp‟

„Bog‟

Water „Fen, Marsh and Swamp‟

Acid grassland

Broadleaved woodland Rough grassland

Any grassland „Built-up Areas and

Gardens‟

Any water „Built-up Areas and

Gardens‟

Any water Any water

Saltwater „Littoral Rock‟

„Littoral Sediment‟